A dual-stream mamba method with propagation environment constraints for sea surface drone multiple-input multiple-output channel prediction

Abstract

The application discloses a double-flow Mamba method with propagation environment constraint for sea surface unmanned aerial vehicle multiple-input multiple-output channel prediction, comprising the following steps: constructing a double-flow Bi-Mamba coding structure: a slow branch extracts long-range evolution characteristics of a sea surface environment and generates a dynamic physical boundary, and a fast branch captures transient detail characteristics of a channel; a physical constraint dynamic mask is generated through a cross-flow gating unit, a physical boundary is updated in real time, channel characteristics are adaptively weighted dimension by dimension, and non-stationary noise without physical basis is inhibited; and a fusion feature is decoded and extrapolated to output a channel prediction result. The application realizes low-delay reasoning by using a linear complexity Mamba architecture, effectively adapts to sea surface environment mutations by using a physical prior constraint, and significantly improves the robustness and physical consistency of channel prediction under complex sea conditions.

Description

Technical Field

[0001] This invention relates to the field of maritime unmanned communication network technology, and specifically to a two-stream Mamba method with propagation environment constraints for multi-input multi-output channel prediction of maritime unmanned aerial vehicles. Background Technology

[0002] With the large-scale deployment of maritime low-altitude unmanned aerial vehicle (UAV) communication networks, building stable, real-time, and reliable maritime MIMO communication links is the core support for realizing services such as marine monitoring, maritime operations, and low-altitude emergency communication. In complex scenarios involving high-speed UAV movement and dynamic changes in sea surface weather, accurate and low-latency acquisition of channel status information is a key prerequisite for dynamic allocation of communication resources, adaptive link control, and ensuring the continuity of UAV communication and operational safety.

[0003] Current prediction methods for time-varying channels are mostly driven by single time-series data, which has significant limitations when facing the complex propagation environment of the sea surface. On the one hand, most prediction models only fit and extrapolate the channel response sequence, detaching themselves from the physical nature of how the sea surface propagation environment affects the signal propagation process. This results in a lack of physical interpretability and a significant increase in prediction errors under conditions of abrupt sea state changes and non-stationary channels. On the other hand, the computational cost of traditional models increases dramatically when facing multi-dimensional parameter predictions. Therefore, in the context of maritime communication, balancing high prediction accuracy with low latency remains a significant challenge in the field.

[0004] In recent years, deep learning methods based on the Transformer architecture have been widely used in time series prediction tasks, capable of capturing long-range dependencies through self-attention mechanisms. However, the inherent quadratic computational complexity of self-attention mechanisms poses a significant challenge when deployed on UAVs with limited computing resources and power consumption, making it difficult to meet the practical requirements of low latency and lightweight inference for maritime UAV communication. Furthermore, existing Transformer-based channel prediction research is still primarily data-driven, generally lacking explicit modeling of physical propagation laws. Environmental and channel features are often simply combined, failing to effectively constrain the channel prediction process through prior environmental information. This results in models exhibiting prediction drift and insufficient generalization ability when facing complex channel environments such as the highly dynamic, non-stationary, and time-varying sea surface, making it difficult to adapt to the stable prediction requirements under complex sea conditions.

[0005] Some existing publicly available technologies have attempted to incorporate physical prior information and adopt a multi-branch parallel modeling approach for time series prediction, improving prediction accuracy by integrating external environmental features. These solutions have shown some effectiveness in general scenarios such as weather forecasting, land communication, and traffic time series prediction. However, research has revealed that this type of general parallel physical prior architecture cannot be directly adapted to the MIMO channel prediction scenario for maritime UAVs, exhibiting significant scenario adaptability defects and technical barriers.

[0006] First, the MIMO channel for UAVs on the sea surface possesses unique dual-time-scale strong coupling characteristics. The sea surface propagation environment parameters exhibit slow, long-period, and trend-based evolution, while the wireless channel, affected by multipath superposition, sea surface scattering, Doppler jitter caused by the high-speed movement of the UAV, and abrupt phase delay changes, exhibits high-frequency, fast-changing, strongly random, and non-stationary instantaneous characteristics. Existing general-purpose parallel branch structures only employ simple feature fusion and equal-weight modeling methods, failing to distinguish the differentiated evolution patterns between the slow-changing trends of the environment and the fast-changing details of the channel in the sea surface scenario, and thus unable to achieve targeted constraint correction of channel dynamics based on physical priors.

[0007] Secondly, existing physical prior parallel modeling schemes are mostly designed for static or slowly changing scenarios, without considering the complex propagation mechanisms unique to marine channels, such as multipath fading, evaporation waveguide propagation, and sea surface clutter interference. The coupling and correlation logic between environmental features and channel features does not conform to the laws of marine radio wave propagation. Blindly applying a general parallel architecture will cause the environmental prior information to become invalid, and it will be unable to effectively suppress instantaneous non-stationary noise in the channel, resulting in a significant decrease in prediction robustness.

[0008] In addition, sudden environmental conditions such as rapid changes in wind speed and large waves frequently occur at sea, causing drastic changes in sea state within a short period. Most existing parallel modeling schemes are trained on stable environmental data, resulting in insufficient dynamic response capabilities and slow convergence speed and poor fault tolerance when facing sudden environmental changes. Furthermore, the high-speed maneuvering of UAVs further exacerbates instantaneous channel fluctuations. Existing schemes struggle to identify channel abrupt changes by incorporating environmental evolution patterns, easily misjudging effective channel changes caused by sudden sea state changes as noise and suppressing them, or retaining random clutter as effective features. With these two problems combined, the prediction error of the model continues to expand in the highly dynamic and rapidly changing real-world maritime conditions, making it difficult to meet the requirements for continuous and stable control of UAV communication links.

[0009] Finally, the mainstream Transformer-type time series models currently exist with quadratic computational complexity, which cannot meet the deployment requirements of lightweight and low latency on the UAV side. While selective state-space models, represented by Mamba and Bi-Mamba, have linear complexity and efficient long sequence modeling capabilities, existing research is designed for general time series prediction tasks. They have not carried out scenario-based structural adaptation and mechanism design for the unique characteristics of the dual time-scale coupling, strong non-stationarity, and strong physical constraints of the MIMO channel of UAVs on the sea surface. Therefore, they cannot achieve deep collaborative and adaptive coupling modeling of physical priors and channel dynamic characteristics.

[0010] In summary, existing general-purpose physical prior parallel modeling techniques, traditional time series prediction models, and lightweight time series networks are all unable to adapt to the unique characteristics of the high dynamics, dual time-scale coupling, and strong non-stationarity of MIMO channels for marine UAVs. The industry urgently needs a channel prediction solution that is suitable for marine communication scenarios, conforms to the physical laws of radio wave propagation over the sea, and balances high accuracy with low computational overhead, in order to solve the technical challenges of poor scenario adaptability, low prediction accuracy, and high deployment difficulty of existing technologies. Summary of the Invention

[0011] To address the technical challenge of existing time-series prediction models being ill-suited to the dual-time-scale coupling and strongly non-stationary propagation characteristics of MIMO channels for maritime UAVs, this invention provides a dual-stream Mamba method with propagation environment constraints for MIMO channel prediction of maritime UAVs. Based on the lightweight time-series modeling capabilities of bidirectional Mamba, a dual-stream decoupled modeling system adapted to the propagation mechanism of radio waves over the sea is constructed. This achieves differentiated, collaborative modeling and adaptive fusion of environmental physical priors and channel dynamic characteristics. While maintaining linear low computational complexity and ensuring the deployability of the UAV at the terminal side, it effectively utilizes the constraints of the maritime propagation environment to correct instantaneous non-stationary fluctuations in the channel. This significantly improves the channel prediction accuracy and generalization capability under complex sea conditions and highly dynamic UAV motion scenarios, providing reliable channel state support for real-time resource scheduling and stable transmission of MIMO communication links for maritime UAVs.

[0012] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:

[0013] A two-stream Mamba method with propagation environment constraints for multiple-input multiple-output channel prediction of maritime unmanned aerial vehicles (UAVs), the method comprising:

[0014] S1 models the sea surface communication channel as the superposition of three key path components: line-of-sight component, sea surface reflection component, and evaporation waveguide reflection component. By integrating sea surface waves, evaporation waveguide effects, and UAV attitude, a non-stationary sea surface communication channel representation with physical consistency is formed, and a low-altitude multi-input multi-output channel model for the sea surface is constructed.

[0015] S2. Based on the channel model, Pierson-Moskowitz power spectrum combined with Beaufort wind partitioning is used to generate multi-level sea state data, and a high-dimensional time series dataset containing multi-dimensional feature variables is obtained through simulation.

[0016] S3. Based on the dataset, the input tensor of the prediction model is deconstructed into two sets of independent parallel input vectors: environmental feature stream and channel state information stream. The continuous historical observation sequence is discretized to establish a mapping relationship between external meteorological parameters and internal channel dynamic response, forming a dual-stream decoupled input structure.

[0017] S4, the environmental feature stream and the channel state information stream are projected to a high-dimensional latent space through a linear embedding layer to generate two sets of corresponding continuous embedding representations;

[0018] S5. A bidirectional Mamba model is constructed using Mamba. The two sets of continuous embedding representations are processed separately based on Mamba blocks to extract cross-variable correlation and intra-variable time dependence, forming an encoding architecture for environmental prior constraints. A slow branch is used to process the embedding representations corresponding to the environmental feature stream to extract the long-range evolution law of the sea surface environment as a physical prior. A fast branch is used to process the embedding representations corresponding to the channel state information stream to model the instantaneous channel response. Both branches use a Bi-Mamba structure to complete bidirectional context feature capture. A cross-stream gating unit is introduced. The environmental prior output of the slow branch is used to generate a dynamic gating mask and adaptively modulate the output of the fast branch to suppress noise fluctuations that do not conform to physical laws, thus completing the fusion of dual-stream features. The fused features are then passed through layer normalization and feedforward neural networks for temporal extrapolation, and then restored to the physical observation space through a linear projection layer to output the final channel prediction result.

[0019] Furthermore, in step S1, the specific process of constructing the low-altitude multi-input multi-output channel model over the sea surface includes:

[0020] For a low-altitude sea surface MIMO communication channel equipped with P transmitting antennas and Q receiving antennas, determine the transceiver antenna pairs. Time-varying frequency response The multipath superposition modeling framework decomposes it into three key path components: line-of-sight, sea surface reflection, and evaporation waveguide reflection.

[0021] For each sub-path under each type of path Define its instantaneous complex amplitude at time t. Phase and transmission delay And construct an exponential term that includes path phase and time delay to characterize the phase evolution characteristics of the path;

[0022] The transceiver attitude rotation matrix is ​​obtained by cascading the basic rotation matrices in the three dimensions of yaw, pitch, and roll. :

[0023] The projections of the radiation patterns of the transmitting and receiving antennas onto the vertical and horizontal polarization planes are coupled with the attitude rotation matrix of the transceiver to form a matrix form of polarization-attitude joint modulation, resulting in a complex gain modulation term that couples the antenna polarization component with the UAV attitude. ;

[0024] By integrating the path components, complex gain modulation terms, and phase delay terms, the transceiver antenna pair is obtained. Time-varying frequency response .

[0025] Step S2 further includes:

[0026] The Pierson-Moskowitz power spectrum combined with the Beaufort wind classification was used to divide the sea surface environment into several typical sea state levels;

[0027] Under each sea state level, different drone receiver movement speeds and communication signal-to-noise ratio conditions are set to form a simulation configuration with multiple scenarios and conditions.

[0028] Based on the low-altitude multi-input multi-output channel model of the sea surface constructed in step S1, channel simulations are carried out under various simulation configurations. High-dimensional time series data containing scene topology, environmental parameters and channel state information are generated through simulation, and a high-fidelity sea surface channel dataset covering multiple sea states and multiple motion states is constructed.

[0029] Step S3 further includes:

[0030] The continuous historical observation sequences in the high-dimensional time series data obtained in step S2 are discretized.

[0031] By combining the core physical parameters affecting the marine electromagnetic propagation environment, a mapping relationship between the external meteorological environment and the internal channel dynamic response is established, and the environmental feature vector at time t is constructed. ;in, For wind speed, For the effective wave height, The height of the evaporation waveguide. This is the sea surface roughness coefficient. The distance between the transmitting and receiving ends, These are the yaw, pitch, and roll angles of the transmitter. For the receiver's yaw angle, pitch angle, and roll angle;

[0032] The MIMO system at time t Mapping the complex channel tensor of each antenna pair to the real domain, we construct the space-time-frequency channel state information vector at time t as follows: ;in, For transceiver antenna pairs The time-varying frequency response between;

[0033] By combining the environmental feature vector and the channel state information vector, two independent and parallel input streams, namely the environmental feature stream and the channel state information stream, are constructed to generate the input tensor of the prediction model. The environmental feature stream and the channel state information stream are respectively composed of the environmental feature vector and the channel state information vector at the corresponding time point in a time sequence.

[0034] Further, in step S4, the input sequences of the environmental feature stream and the channel state information stream are processed in batches and mapped to the high-dimensional latent space through independent linear projection layers to obtain two sets of corresponding embedding representations, denoted as environmental feature embedding and channel state embedding, respectively; wherein, the dimension of each set of embedding representations is B×L×D, where B is the batch size, L is the sequence length, and D is the hidden dimension.

[0035] Further, in step S5, the prediction model sequentially includes a lexicalization layer, a two-stream Mamba coding layer, a cross-stream gating unit layer, and a linear decoding layer;

[0036] The lexicalization layer performs lexicalization on the two sets of consecutive embedding representations obtained in step S4 to obtain the environmental feature stream lexical sequence. and channel state information stream lexical sequence Among them, environmental feature flow word sequence Temporal characteristics used to characterize environmental factors including sea surface weather and UAV attitude; channel state information stream term sequence Used to characterize the dynamic changes in the space-time-frequency response of a MIMO channel;

[0037] The dual-stream Mamba coding layer includes a slow branch and a fast branch; wherein, the slow branch adopts a Bi-Mamba structure to encode the context feature stream word sequence. Temporal modeling is performed to extract the long-term evolution trends of environmental parameters, including sea surface wind speed, wave height, evaporation waveguide height, and UAV attitude, to characterize the continuous changes in the sea surface propagation environment and form environmental constraint features with physical prior attributes. The fast branch uses a Bi-Mamba structure to stream the term sequence of channel state information. Time-series modeling is performed to capture the details of instantaneous channel abrupt changes caused by multipath superposition, Doppler jitter, and phase delay fluctuations, thus characterizing the short-time rapid dynamic response of non-stationary sea surface channels and forming detailed channel state features. ;

[0038] The cross-current gating unit layer will define the environmental constraint characteristics of the slow branch output. A dynamic gating mask is generated through a linear layer, and the mask is fine-tuned element-by-element based on the physical boundary of sea surface radio wave propagation: the mask value is increased for channel mutation dimensions consistent with the long-term evolution trend of the environment to retain information on compliant environmental mutations; the mask value is decreased for noise fluctuation dimensions that deviate from physical laws to suppress invalid anomalous components; and this mask is used to extract detailed channel state features from the fast tributary output. Modulation is performed, and then the two-stream features are fused through residual connection and normalization operations to obtain the fused features. This allows the channel characteristics to carry constraint information about the sea surface propagation environment;

[0039] The linear decoding layer will input fused features After time-domain extrapolation using layer normalization and feedforward neural network, the data is then restored to the original channel physical observation space through a linear projection layer, outputting the channel state prediction results for the MIMO of the UAV at future time.

[0040] Furthermore, the cross-flow gating unit layer first carries the environmental constraint characteristics of the long-range evolution law of the sea surface environment. The input is fed into a linear transform layer for feature dimension alignment. A dynamic gating mask is generated through the linear layer and the Sigmoid activation function. This mask adaptively adjusts the weight allocation of channel features based on the changing trends of environmental parameters, including sea surface wind speed, wave height, evaporation waveguide height, and UAV attitude. Then, the dynamic gating mask is used to refine the channel state details output by the fast branch, including multipath superposition, Doppler jitter, and phase delay fluctuations. Element-wise Hadamard product modulation is performed to obtain detailed channel state features. The various dimensions are adaptively weighted to suppress channel noise and non-stationary fluctuations that do not conform to the physical laws of the environment; finally, the modulated channel features are compared with the original channel state details. Residual connections are performed, and then layer normalization is used to complete the fusion of dual-stream features, resulting in fused features that take into account both long-term environmental constraints and instantaneous channel details.

[0041] Furthermore, the process of adaptively adjusting the weight allocation of channel features based on the changing trends of environmental parameters, including sea surface wind speed, wave height, evaporation waveguide height, and UAV attitude, includes the following steps:

[0042] Environmental constraints that carry the long-term evolution law of the marine environment The input linear transform layer is used for dimension alignment to generate detailed features of the fast tributary channel state. Dimensional matching mask generation features;

[0043] A dynamic gated mask is generated by using a linear layer and a Sigmoid activation function. The mask has a value range of [0,1], and each mask element corresponds to one dimension of the channel state detail features.

[0044] The pre-trained physical agent network is invoked, and environmental parameters including sea surface wind speed, wave height, evaporation waveguide height, and UAV attitude are used as input to output the dynamic physical boundary corresponding to the propagation of sea surface radio waves in real time. The dynamic physical boundary includes the multipath delay interval, the maximum Doppler frequency shift threshold, and the upper and lower limits of the signal energy envelope.

[0045] Based on the dynamic physical boundary, the channel state details are described. Compliance is determined on a dimensional basis. Dimensions whose feature values ​​fall within the physical boundary range are determined to be compliant channel dimensions, while dimensions whose feature values ​​exceed the physical boundary or exhibit irregular instantaneous jumps are determined to be noise dimensions.

[0046] Based on the dimensionality determination results, the dynamic gating mask is fine-tuned element by element: the mask elements corresponding to the compliant channel dimension are numerically increased to amplify the effective feature weights, while the mask elements corresponding to the noise dimension are numerically decreased to weaken the abnormal fluctuation weights, and the mask values ​​remain within the [0,1] range after correction; then the fine-tuned mask is used to analyze the channel state detail features. Element-wise Hadamard product modulation is performed to complete the weight allocation of each channel dimension, thereby suppressing the portion of channel noise and non-stationary fluctuations that does not conform to the physical laws of the environment.

[0047] Further, in step S5, the prediction model uses a joint loss function based on physical environment boundary constraints for model training and parameter updates. This joint loss function is a weighted sum of the channel prediction mean square error term and the physical environment boundary constraint regularization term.

[0048] ;

[0049] in, For the total loss, This is the balance coefficient; This is the channel prediction mean square error term, used to constrain the deviation between the channel state prediction results output by the model and the actual channel state; The physical environment boundary constraint regularization term utilizes a pre-trained physical proxy network and channel feature extraction operators to construct dynamic physical boundary conditions, penalizing abnormal results in the channel state information flow prediction results that violate the physical constraints of sea surface channel propagation.

[0050] Furthermore, the physical agent network is a multilayer perceptron architecture, and pre-training is completed through offline supervised learning before the prediction model is trained; in step S5, the physical agent network takes the environmental feature stream as input and infers a multi-dimensional physical evolution boundary threshold vector, which includes the upper bound of the maximum Doppler frequency shift, the maximum multipath delay spread limit, and the energy envelope limit.

[0051] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0052] First, the dual-stream Mamba method with propagation environment constraints for MIMO channel prediction of maritime unmanned aerial vehicles (UAVs) of this invention comprehensively integrates various typical marine propagation factors such as sea surface wind speed, significant wave height, evaporation duct height, and UAV attitude perturbations. This fully covers the physical causes affecting the evolution of MIMO channels for maritime UAVs, avoiding the modeling distortion problem caused by traditional data-driven models that ignore environmental physical constraints and rely solely on channel sequence fitting. By decoupling the original time-series data into two types of input modes with different physical attributes—environmental feature stream and channel state information stream—it achieves independent representation of the slowly varying physical background and the rapidly changing channel dynamics. This effectively solves the deficiency of existing single-input, general parallel structures that cannot distinguish the dual-timescale characteristics of the sea surface, providing a reliable data and representation foundation for subsequent high-precision, interpretable channel prediction modeling.

[0053] Secondly, the dual-stream Mamba method with propagation environment constraints for multi-input multi-output channel prediction of maritime UAVs in this invention addresses the unique dual-scale coupling characteristics of the maritime channel: long-range slow-change environment and instantaneous rapid-change channel. Leveraging the advantages of bidirectional Mamba linear time-series modeling, it captures the environmental trend evolution and high-frequency dynamic details of the channel through a differentiated dual-branch modeling approach. This overcomes the problems of existing general-purpose Mamba models and Transformer models being unable to adapt to the multi-scale physical characteristics of the sea surface and lacking specific modeling targeting. Simultaneously, the overall architecture maintains linear computational complexity independent of sequence length, significantly reducing computational latency and resource consumption compared to traditional quadratic complexity time-series models, effectively meeting the deployment requirements of low latency, lightweight design, and high real-time performance in maritime UAV communication scenarios.

[0054] Third, the dual-stream Mamba method with propagation environment constraints for multi-input multi-output channel prediction of maritime unmanned aerial vehicles (UAVs) of this invention addresses the problems of loose fusion of environmental and channel features and failure of physical prior constraints in existing parallel modeling schemes. It utilizes stable environmental physical prior information learned from the slow branch to dynamically modulate the evolution process of instantaneous channel features in the fast branch, achieving adaptive constraint and correction of macroscopic environmental laws on microscopic channel dynamics. This cross-stream physical guidance mechanism effectively suppresses invalid noise fluctuations caused by sea surface clutter, random multipath, and Doppler jitter, ensuring that the channel prediction results always conform to the physical laws of sea surface radio wave propagation, significantly improving the model's prediction stability and generalization ability under highly dynamic, non-stationary, and unknown sea conditions. Attached Figure Description

[0055] Figure 1 This is a general block diagram of a dual-stream Mamba channel prediction architecture;

[0056] Figure 2 Here is a diagram of the Mamba block structure;

[0057] Figure 3 Comparison of model prediction results and prediction errors for four sea state scenarios and with or without environmental feature flows;

[0058] Figure 4 The graph shows a performance comparison between the model of this invention and other baseline models in four sea state scenarios, in terms of normalized mean square error, training time, and inference time.

[0059] Figure 5 This is a flowchart of the two-stream Mamba method with propagation environment constraints for multi-input multi-output channel prediction of maritime unmanned aerial vehicles (UAVs) according to the present invention. Detailed Implementation

[0060] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0061] This invention discloses a two-stream Mamba method with propagation environment constraints for multiple-input multiple-output channel prediction of maritime unmanned aerial vehicles (UAVs), the method comprising:

[0062] S1 models the sea surface communication channel as the superposition of three key path components: line-of-sight component, sea surface reflection component, and evaporation waveguide reflection component. By integrating sea surface waves, evaporation waveguide effects, and UAV attitude, a non-stationary sea surface communication channel representation with physical consistency is formed, and a low-altitude multi-input multi-output channel model for the sea surface is constructed.

[0063] S2. Based on the channel model, Pierson-Moskowitz power spectrum combined with Beaufort wind partitioning is used to generate multi-level sea state data, and a high-dimensional time series dataset containing multi-dimensional feature variables is obtained through simulation.

[0064] S3. Based on the dataset, the input tensor of the prediction model is deconstructed into two sets of independent parallel input vectors: environmental feature stream and channel state information stream. The continuous historical observation sequence is discretized to establish a mapping relationship between external meteorological parameters and internal channel dynamic response, forming a dual-stream decoupled input structure.

[0065] S4, the environmental feature stream and the channel state information stream are projected to a high-dimensional latent space through a linear embedding layer to generate two sets of corresponding continuous embedding representations;

[0066] S5. A bidirectional Mamba model is constructed using Mamba. The two sets of continuous embedding representations are processed separately based on Mamba blocks to extract cross-variable correlation and intra-variable time dependence, forming an encoding architecture for environmental prior constraints. A slow branch is used to process the embedding representations corresponding to the environmental feature stream to extract the long-range evolution law of the sea surface environment as a physical prior. A fast branch is used to process the embedding representations corresponding to the channel state information stream to model the instantaneous channel response. Both branches use a Bi-Mamba structure to complete bidirectional context feature capture. A cross-stream gating unit is introduced. The environmental prior output of the slow branch is used to generate a dynamic gating mask and adaptively modulate the output of the fast branch to suppress noise fluctuations that do not conform to physical laws, thus completing the fusion of dual-stream features. The fused features are then passed through layer normalization and feedforward neural networks for temporal extrapolation, and then restored to the physical observation space through a linear projection layer to output the final channel prediction result.

[0067] Maritime communication environments possess unique characteristics, with channel properties influenced by both marine meteorological conditions and the dynamic characteristics of both platforms. At the physical propagation level, the maritime channel exhibits a complex multipath structure due to environmental factors. At the transceiver platform level, attitude changes in the UAV platform significantly alter antenna gain and directivity. Based on this, this invention constructs a low-altitude maritime channel model that comprehensively considers both the maritime environment and platform dynamics, aiming to achieve a high-precision characterization of the complex maritime propagation environment, thereby providing accurate data representation and physical support for subsequent prediction algorithms.

[0068] This invention models the sea surface communication channel as a superposition of three critical path components: the line-of-sight (LoS) component, the sea surface reflection (SSR) component, and the evaporation duct reflection (EDR) component. For a low-altitude sea surface MIMO communication channel equipped with P transmit antennas and Q receive antennas, the transmit and receive antenna pairs... Time-varying frequency response It can be modeled as:

[0069] ;

[0070] in, and , For the set of propagation paths, This represents the total number of corresponding paths. This indicates the complex gain modulation of the transmit and receive antenna patterns on each path polarization component. , and They represent the first The path is The instantaneous complex amplitude, phase, and propagation delay at each moment characterize the phase evolution characteristics of the path. Considering the influence of UAV attitude changes and antenna directivity on the signal amplitude and phase characteristics, the complex gain modulation term of the transmitting and receiving antenna patterns coupled with the UAV attitude for each path polarization component is calculated. It can be further expressed as

[0071] ;

[0072] in, and The first The emission angle of the transmitter path under the transmit / receive polarization p,q, and the incident angle of the receiver's line-of-sight path. and Let represent the projections of the transmitting and receiving antenna radiation patterns onto the vertical and horizontal polarization planes, respectively. Couple these projections with the transceiver attitude rotation matrix to form a matrix form for polarization-attitude joint modulation. This yields the complex gain expression for the coupling between the antenna polarization component and the UAV attitude, and the transceiver attitude rotation matrix. It is obtained by cascading the basic rotation matrices of yaw, pitch, and roll, and is represented as:

[0073] ;

[0074] in, , and They represent the yaw angles respectively. Pitch angle and roll angle The fundamental rotation matrix in the direction.

[0075] Due to the scarcity and high cost of experimental data on maritime communication, this invention utilizes Quadriga to construct a maritime channel simulation platform based on the established maritime channel model, generating a large amount of high-fidelity channel data for model training. Specifically, the process includes the following steps: First, the maritime environment is divided into multiple typical sea state levels using Pierson-Moskowitz power spectrum analysis combined with Beaufort wind force partitioning. Second, under each sea state level, different UAV receiver movement speeds and communication signal-to-noise ratio conditions are set to form multi-scenario, multi-condition simulation configurations. Third, based on the low-altitude multi-input multi-output (MIMO) channel model constructed in step S1, channel simulations are performed under each simulation configuration. The simulation generates high-dimensional time-series data containing scene topology, environmental parameters, and channel state information, constructing a high-fidelity maritime channel dataset covering multiple sea states and motion states.

[0076] To cover various maritime propagation environments and generate ergonomic sea surface channel data, this embodiment utilizes the Pierson-Moskowitz power spectrum and combines it with the Beaufort wind classification method to divide sea state (SS) into four representative levels: SS2, SS6, SS8, and SS10. It should be understood that sea state levels are classified according to actual needs and are not limited to this embodiment.

[0077] Multi-condition experiments were conducted at each sea state level, covering different receiver moving speeds (0–20 m / s) and signal-to-noise ratios (0–30 dB). Each simulation lasted 10 seconds, generating a high-dimensional time series of length 10,001. The data incorporated scene geometry, environmental parameters, and channel state information, with each data sample including 523 multi-dimensional feature variables. The simulation configuration used a 16×16 MIMO antenna system with a center frequency of 28 GHz and a system bandwidth of 50 MHz. Specific simulation conditions are shown in Table 1.

[0078] Table 1 Simulation parameters and configurations

[0079]

[0080] In order to fully exploit the prior knowledge in the aforementioned channel model, this embodiment deconstructs the input tensor of the prediction model into an environmental feature flow and a channel state information flow, aiming to establish a mapping relationship between external meteorological parameters and internal channel dynamic response, thereby enhancing the physical interpretability and prediction accuracy of the prediction model for complex sea surface scenarios.

[0081] The continuous historical observation sequence is discretized, and two sets of independent and parallel input vectors are constructed. in, and These are the environmental feature stream and the channel state information stream, respectively. It incorporates the core physical parameters affecting the marine electromagnetic propagation environment and defines... The vector representation of the environmental feature flow at time step is as follows:

[0082] ;

[0083] This includes wind speed Significant wave height Evaporation waveguide height Sea surface roughness coefficient Distance between transceiver ends and the attitude angle of the transmitting and receiving ends , By incorporating these long-range physical contexts, the model can use them as predictive constraints to improve prediction accuracy. Channel state information flow. This refers to channel state information influenced by the propagation environment. To facilitate neural network processing, [the following is omitted as it is not directly related to the preceding text]. Time-based MIMO system The complex channel tensor of each antenna pair is mapped to the real domain, and the space-time-frequency channel state information is defined as follows:

[0084] ;

[0085] In the formula, For transceiver antenna pairs The time-varying frequency response between.

[0086] Based on this, this embodiment models the sea surface UAV channel prediction task as a multivariate time series prediction task, the objective of which is to predict the channel based on known data. Historical samples Accurately predict the future Output at time 1 .

[0087] In low-altitude sea-surface mobile communication scenarios, channel evolution is subject to complex joint constraints from multi-dimensional propagation environment parameters. However, existing channel prediction studies largely rely on single-dimensional historical channel state information, neglecting the intrinsic evolutionary mechanism between the physical environment and channel response. This black-box modeling paradigm limits the model's scenario generalization ability and leads to a lack of physical interpretability. To address these challenges, this embodiment proposes a two-stream Mamba prediction framework. Figure 1 As shown, this framework extracts intervariate correlations and intravariate time dependencies based on Mamba blocks. By simultaneously introducing environmental feature streams and channel state information streams, it constructs a predictive architecture with environmental prior constraints. Subsequently, a cross-stream gating unit is designed to inject constraints from the sea surface environment onto the channel evolution state.

[0088] The core principle of Mamba lies in establishing a non-linear mapping between system inputs and outputs by introducing hidden states. Figure 1 The algorithmic principle of the basic Mamba block is presented. The Mamba block constructs an efficient sequence modeling unit by integrating linear projection, convolution operations, and gating mechanisms. The input sequence is processed in batches and then mapped to a high-dimensional space via linear projection to obtain... Where B is the batch size, and L and D are the sequence length and hidden dimension, respectively. First, a linear mapping layer expands the hidden dimension to ED, where... The block expansion factor is then divided into main branches. With gated branches In the main branch, the model first extracts local temporal correlation features using one-dimensional convolution, and then processes these features using the SiLU activation function to obtain enhanced features. Subsequently, Applying the aforementioned selective state-space framework, the model dynamically generates discretized parameters based on the input content to realize the system's state mapping, generating an output state that captures long-range dependencies. This selective mechanism endows Mamba with an ability to perceive effective information among multiple variables, similar to an attention mechanism. (Final state) The elements are multiplied one-to-one with the gated flow after activation, and the dimensions are reconstructed back using a linear transformation. The final output tensor is obtained. .

[0089] Based on this, this embodiment first processes the input sequences of the environmental feature stream and the channel state information stream in batches, and then maps them to a high-dimensional hidden space through independent linear projection layers to obtain two sets of corresponding embedding representations, denoted as environmental feature embedding and channel state embedding, respectively; wherein, the dimension of each set of embedding representations is B×L×D, where B is the batch size, L is the sequence length, and D is the hidden dimension.

[0090] A prediction model based on Bi-Mamba is constructed, which sequentially includes a lexicalization layer, a dual-stream Mamba coding layer, a cross-stream gating unit layer, and a linear decoding layer. The lexicalization layer performs lexicalization on the two sets of consecutive embedding representations to obtain the lexical sequence corresponding to the environmental feature stream. The word sequence corresponding to the channel state information stream Among them, the word sequence Temporal features used to characterize environmental factors including sea surface weather and UAV attitude, word sequence The dual-stream Mamba coding layer is used to characterize the dynamic changes in the space-time-frequency response of the MIMO channel; it includes a slow branch and a fast branch; wherein, the slow branch uses a Bi-Mamba structure to encode the environmental feature stream term sequence. Temporal modeling is performed to extract the long-term evolution trends of slowly varying environmental parameters, including sea surface wind speed, wave height, evaporation waveguide, and UAV attitude, thus characterizing the continuous changes in the sea surface propagation environment and forming environmental constraint features with physical prior attributes. The fast branch uses a Bi-Mamba structure to stream the term sequence of channel state information. Time-series modeling is performed to capture the details of instantaneous channel abrupt changes caused by multipath superposition, Doppler jitter, and phase delay fluctuations, thus characterizing the short-time rapid dynamic response of non-stationary sea surface channels and forming detailed channel state features. The cross-current gating unit layer will define the environmental constraint characteristics of the slow branch output. A dynamic gating mask is generated through a linear layer, and the mask is fine-tuned element-by-element based on the physical boundary of sea surface radio wave propagation: the mask value is increased for channel mutation dimensions consistent with the long-term evolution trend of the environment to retain information on compliant environmental mutations; the mask value is decreased for noise fluctuation dimensions that deviate from physical laws to suppress invalid anomalous components; and this mask is used to extract detailed channel state features from the fast tributary output. Modulation is performed, and then the two-stream features are fused through residual connection and normalization operations to obtain the fused features. This allows the channel features to carry constraint information about the sea surface propagation environment; the linear decoding layer then fuses the input features. After time-domain extrapolation using layer normalization and feedforward neural network, the data is then restored to the original channel physical observation space through a linear projection layer, outputting the channel state prediction results for the MIMO of the UAV at future time.

[0091] To reduce inference latency and enhance real-time performance, this example demonstrates a multi-parameter prediction module for the sea surface channel based on Mamba. To adapt the raw channel observation data to the input requirements of the Mamba block, a linear embedding layer is first introduced to process the input sequence. Feature mapping is performed. By projecting the original observation space onto a high-dimensional latent space, a continuous embedding representation that the model can understand is generated. This process can be represented as:

[0092] ;

[0093] in, The input sequence contains both environmental feature stream and channel state information stream. For the lexicalized feature representation, B, L, and V represent the batch size, sequence length, and number of variables, respectively, and D represents the hidden dimension. This is a layer normalization operation.

[0094] The unidirectional causal nature of the standard Mamba architecture limits its ability to perceive the global context, enabling it to integrate only historical features prior to the current time step. To overcome this physical limitation, this paper introduces a bidirectional Mamba (Bi-Mamba) structure. Based on the Mamba structure, through the parallel structure of forward and reverse Mamba operators, this architecture can break through the unidirectional time-domain constraint, endowing the model with the ability to model the correlation between variables and the time dependence within variables. This bidirectional scan operation can be expressed as...

[0095] ;

[0096] in, For Bi-Mamba operators, and These are forward and backward Mamba blocks, respectively. This indicates a feature fusion operation.

[0097] To enhance the model's ability to handle complex maritime communication scenarios, a dual-stream architecture is proposed. This architecture uses a slow branch and a fast branch to handle environmental priors and channel responses respectively, aiming to achieve strong constraints from the physical environment on channel prediction behavior. The slow branch aims to extract long-range time-series evolution patterns from meteorological environmental parameters, providing physical constraints for channel response prediction in the fast branch. Furthermore, both branches employ a Bi-Mamba structure, enabling the model to deeply capture bidirectional contextual features and dynamically inject physical environmental constraints. The processing procedures of the slow and fast branches can be represented as follows:

[0098] ;

[0099] ;

[0100] in, and These represent the embedded environmental feature stream and channel state information stream inputs, respectively. and These are the Bi-Mamba operators corresponding to the two branches. and Do not use two independent sets of learnable parameters for each branch.

[0101] Subsequently, this embodiment further introduces a cross-stream gated unit (CSGU) to achieve feature fusion of dual-stream Mamba, thereby realizing the framework design for channel response prediction constrained by physical environment parameters. Specifically, firstly, environmental constraint features carrying the long-range evolution law of the sea surface environment are... The input is fed into a linear transformation layer for feature dimension alignment, and a dynamic gating mask is generated through the linear layer and the Sigmoid activation function. This mask adaptively adjusts the weighting of channel features based on the changing trends of environmental parameters, including sea surface wind speed, wave height, evaporation waveguide height, and UAV attitude, and applies these weights to the fast branch output. This enhances the channel state details of the fast branch output, including multipath superposition, Doppler jitter, and phase delay fluctuations. Element-wise Hadamard product modulation is performed to obtain detailed channel state features. The various dimensions are adaptively weighted to suppress channel noise and non-stationary fluctuations that do not conform to the physical laws of the environment. Finally, the modulated channel features are compared with the original channel state details. Residual connections are performed, and then layer normalization is used to complete the fusion of dual-stream features, resulting in fused features that take into account both long-term environmental constraints and instantaneous channel details. This process can be represented as:

[0102] ;

[0103] ;

[0104] in, This represents the Hadamard product. Subsequently, the fused features are extrapolated in the temporal domain through a decoding module consisting of layer normalization (LN) and a feedforward neural network (FFN). Finally, a linear projection layer restores the abstract latent features to the physical observation space, generating the future... Channel prediction results at each time point :

[0105] ;

[0106] .

[0107] This embodiment achieves physical constraint weighting and noise suppression of non-stationary channel characteristics on the sea surface through a physically constrained dynamic gating mask. The environmental constraint features extracted by the slow branch Bi-Mamba will synchronously reflect the evolution trend of the sea state. The pre-trained physical proxy network updates the dynamic physical boundary of sea surface radio wave propagation in real time accordingly. Based on the updated physical boundary, the dynamic gating mask performs a dimensional compliance judgment on the channel state details, distinguishes between effective channel fluctuations caused by real environmental changes and abnormal noise spikes without physical basis, and allocates differentiated weights to each channel dimension accordingly.

[0108] In a preferred embodiment, to further improve the physical constraints and environmental adaptation accuracy of mask weight allocation, this invention also proposes a dynamic gating mask generation and fine-tuning method that combines the physical boundaries of sea surface radio wave propagation. The specific steps are as follows:

[0109] First, environmental constraints that carry the long-term evolution patterns of the marine environment will be considered. The input linear transform layer is used for dimension alignment to generate detailed features of the fast tributary channel state. The mask generation features are then matched for dimension. Next, these mask generation features are input into a linear layer, and a dynamic gated mask is generated using a Sigmoid activation function. The mask values ​​range from [0,1], and each mask element corresponds to one dimension of the channel state detail features. Then, a pre-trained physical proxy network is invoked, taking environmental parameters including sea surface wind speed, wave height, evaporation waveguide height, and UAV attitude as input, to output the dynamic physical boundary corresponding to sea surface radio wave propagation in real time. This dynamic physical boundary includes the multipath delay interval, the maximum Doppler frequency shift threshold, and the upper and lower limits of the signal energy envelope. Subsequently, the channel state detail features are processed based on the aforementioned dynamic physical boundary. Compliance is assessed dimension by dimension. Dimensions whose feature values ​​fall within the physical boundary are classified as compliant channel dimensions, while dimensions whose feature values ​​exceed the physical boundary or exhibit irregular, transient jumps are classified as noise dimensions. Finally, the dynamic gating mask is fine-tuned element by element based on the dimension assessment results: the mask elements corresponding to compliant channel dimensions are numerically increased to amplify the effective feature weights, while the mask elements corresponding to noise dimensions are numerically decreased to weaken abnormal fluctuation weights. The corrected mask values ​​remain within the [0,1] range. The fine-tuned mask is then used to assess the channel state details. Element-wise Hadamard product modulation is performed to complete the weight allocation of each channel dimension, thereby suppressing the portion of channel noise and non-stationary fluctuations that does not conform to the physical laws of the environment.

[0110] The adaptive weight allocation mechanism of this invention is particularly suitable for marine scenarios with high frequency of environmental changes and complex and variable sea conditions. It can accurately identify and differentiate channel characteristics based on the physical boundary of sea surface radio wave propagation. When a sudden increase in sea surface wind speed or wave height leads to enhanced multipath scattering, the detailed characteristics of the channel state will simultaneously exhibit effective fluctuations caused by changes in the actual propagation path, as well as invalid spikes generated by sea surface clutter and system disturbances. Traditional general-purpose masks or fixed-weight schemes, lacking constraints from environmental physical boundaries, cannot distinguish the essential differences between the two types of fluctuations: if equal weighting is used, noise spikes will be amplified, leading to drastic jumps in prediction results; if sudden fluctuations are excessively suppressed, effective channel changes caused by sea state changes are easily misjudged as noise, causing prediction deviation drift and making it difficult to adapt to the dynamic evolution characteristics of non-stationary sea surface channels. The dynamic gating mask in this embodiment can automatically increase the weight of the effective channel dimension that is caused by environmental changes and conforms to the physical laws of radio wave propagation, based on the real-time updated physical constraints of sea surface propagation, while reducing the weight of abnormal fluctuation dimension without physical basis. This achieves differentiated processing of effective channel changes and invalid noise interference, avoiding misjudgment and distortion of traditional fixed weights or general masks in environmental change scenarios, and significantly improving the robustness and physical consistency of channel prediction under complex sea conditions.

[0111] The algorithm framework proposed in this embodiment is implemented in PyTorch and uses the Adam optimizer for parameter updates. During the model training phase, to ensure the prediction model not only minimizes numerical prediction bias but also incorporates physical environment constraints, a joint loss function for physical environment boundary constraints is constructed based on a pre-trained physical agent network. The total loss of this joint loss function is... Mean square error term of channel prediction Physical environment boundary constraint regularization term Composition, defined as follows:

[0112] ;

[0113] in, This is the balance coefficient used to balance the weights of the two loss terms. Used to minimize the predicted value Compared with the true value The mean square error of the deviation between them is defined as follows:

[0114] ;

[0115] in, express Norm.

[0116] To incorporate prior knowledge into the dual-stream branch prediction phase and prevent the model from outputting predictions that violate the constraints of the physical propagation environment under complex sea conditions, a physical environment boundary constraint regularization term is introduced into the model training phase based on the mean square error. The definition is as follows:

[0117] ;

[0118] In the formula, This is a one-sided truncation penalty function that takes the maximum value of each element. For physical agent networks, a lightweight multilayer perceptron architecture is adopted, which is based on environmental feature flow vectors. The output is a multi-dimensional channel physical boundary vector used to constrain the prediction results of the prediction model. Pre-training is performed offline under supervised learning before the prediction model is trained. Using environmental feature streams as input, the inference output contains a multi-dimensional physical evolution boundary threshold vector, including the upper bound of the maximum Doppler frequency shift, the maximum multipath delay spread bound, and the energy envelope bound. When the instantaneous channel features output by the prediction model do not exceed the physical evolution boundaries given by the physical proxy network, this result is a zero vector, and no penalty is applied. When the predicted instantaneous features undergo a physically inconsistency that violates the physical rules, a difference vector greater than zero is generated, resulting in a quadratic penalty gradient that is backpropagated to the prediction model, ensuring that the final prediction result conforms to the macroscopic physical boundary conditions determined by the sea surface propagation environment. This physical proxy network is defined as follows:

[0119] ;

[0120] in, This is the boundary of the Doppler frequency shift. For multipath delay extension boundary, and This represents the upper and lower boundaries of the spatial energy envelope. Detailed calculation procedures are given in Algorithm 1.

[0121]

[0122] Example

[0123] To cover various maritime propagation environments and generate ergonomic sea surface channel data, this example utilizes the Pierson-Moskowitz power spectrum and combines it with the Beaufort wind classification method to divide Sea State (SS) into four representative levels: SS2, SS6, SS8, and SS10. Multi-condition experiments were conducted at each SS level, covering different receiver moving speeds (0–20 m / s) and signal-to-noise ratios (0–30 dB). Each simulation lasted 10 seconds, generating a high-dimensional time series of length 10,001. The data integrates scene geometry, environmental parameters, and channel state information, with each data sample including 523 multi-dimensional feature variables. The simulation configuration used a 16×16 MIMO antenna system with a center frequency of 28 GHz and a system bandwidth of 50 MHz. Specific simulation conditions are shown in Table 1.

[0124] The algorithm framework proposed in this example is implemented in PyTorch and uses the Adam optimizer for parameter updates. During training, the BatchSize is set to 16, and the initial learning rate is set to... The learning rate employs a round-by-round decay strategy, and its update formula is defined as follows: ,in The initial learning rate is set to 256. Regarding network structure parameters, both the hidden layer dimension and the feedforward network dimension are set to 256.

[0125] To comprehensively evaluate the channel prediction performance of the proposed model in maritime UAV communication scenarios, this paper presents detailed comparative results from two dimensions: cross-sea state quantification assessment and time-domain non-stationary conditions. This embodiment selects several advanced time series prediction architectures and LLM4CP as benchmark models for comparative analysis.

[0126] Figure 3 This paper presents the channel state information (CSO) prediction results of the proposed dual-stream Mamba model under four typical sea states, with and without the incorporation of environmental feature flow constraints. The normalized mean square error (MSE) results for each sea state are also provided. The results show that by integrating prior environmental knowledge, the model's prediction curves better fit the actual values ​​in all sea states, improving the model accuracy by 50.11%. This verifies the constraining effect of environmental flow on the prediction of the CSO sequence. Specifically, the slow branch effectively extracts long-range constraint features from the environmental parameters, while the fast branch accurately captures the complex fluctuations in the CSO sequence. Therefore, this feature fusion strategy combining environmental flow information effectively improves the model's prediction accuracy in non-stationary sea surface environments.

[0127] Figure 4A multi-dimensional bubble diagram is presented to compare the overall performance of the proposed model with other models in terms of prediction accuracy, training efficiency, and inference latency. The four sub-diagrams in the diagram correspond to test scenarios under different sea conditions. The model located in the lower left corner with the smallest bubble diameter represents the optimal trade-off between accuracy, convergence speed, and real-time performance. Experimental results show that, limited by the computational complexity of the global attention mechanism, traditional Transformer and Informer models exhibit higher training time and prediction errors, generally distributed in the upper right quadrant. In contrast, the proposed dual-stream Mamba model consistently occupies the lower left quadrant under all sea conditions. Thanks to the linear computational complexity and efficient autoregressive inference characteristics of the state-space model, the model achieves high training efficiency (approximately 15-18 ms / iter) while significantly reducing prediction error, and maintains low inference latency (approximately 3-5 ms). Compared to the traditional channel prediction model LLM4CP, its inference latency is 25.86% lower.

[0128] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0129] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A two-stream Mamba method with propagation environment constraints for multiple-input multiple-output channel prediction of maritime unmanned aerial vehicles (UAVs), characterized in that, The method includes: S1 models the sea surface communication channel as the superposition of three key path components: line-of-sight component, sea surface reflection component, and evaporation waveguide reflection component. By integrating sea surface waves, evaporation waveguide effects, and UAV attitude, a non-stationary sea surface communication channel representation with physical consistency is formed, and a low-altitude multi-input multi-output channel model for the sea surface is constructed. S2. Based on the channel model, Pierson-Moskowitz power spectrum combined with Beaufort wind partitioning is used to generate multi-level sea state data, and a high-dimensional time series dataset containing multi-dimensional feature variables is obtained through simulation. S3. Based on the dataset, the prediction model input tensor is deconstructed into two sets of independent parallel input vectors: environmental feature stream and channel state information stream. The continuous historical observation sequence is discretized to establish a mapping relationship between external meteorological parameters and internal channel dynamic response, forming a dual-stream decoupled input structure. S4, the environmental feature stream and the channel state information stream are projected to a high-dimensional latent space through a linear embedding layer to generate two sets of corresponding continuous embedding representations; S5. A prediction model is constructed using bidirectional Mamba. Based on Mamba blocks, the two sets of continuous embedding representations are processed separately to extract cross-variable correlation and intra-variable time dependence, forming an encoding architecture for environmental prior constraints. A slow branch is used to process the embedding representations corresponding to the environmental feature stream to extract the long-range evolution law of the sea surface environment as a physical prior. A fast branch is used to process the embedding representations corresponding to the channel state information stream to model the instantaneous channel response. Both branches use a Bi-Mamba structure to complete bidirectional context feature capture. A cross-stream gating unit is introduced. The environmental prior output of the slow branch is used to generate a dynamic gating mask and adaptively modulate the output of the fast branch to suppress noise fluctuations that do not conform to physical laws, thus completing the fusion of dual-stream features. The fused features are then passed through layer normalization and feedforward neural networks for temporal extrapolation, and then restored to the physical observation space through a linear projection layer to output the final channel prediction result. In step S1, the specific process of constructing the low-altitude multi-input multi-output channel model over the sea includes: For a low-altitude sea surface MIMO communication channel equipped with P transmit antennas and Q receive antennas, determine the transmit and receive antenna pairs. Time-varying frequency response The multipath superposition modeling framework decomposes it into three key path components: line-of-sight, sea surface reflection, and evaporation waveguide reflection. For each sub-path under each type of path Define its instantaneous complex amplitude at time t. Phase and transmission delay And construct an exponential term that includes path phase and time delay to characterize the phase evolution characteristics of the path; The transceiver attitude rotation matrix is ​​obtained by cascading the basic rotation matrices in the three dimensions of yaw, pitch, and roll. : The projections of the radiation patterns of the transmitting and receiving antennas onto the vertical and horizontal polarization planes are coupled with the attitude rotation matrix of the transceiver to form a matrix form of polarization-attitude joint modulation, resulting in a complex gain modulation term that couples the antenna polarization component with the UAV attitude. ; By integrating the path components, complex gain modulation terms, and phase delay terms, the transceiver antenna pair is obtained. Time-varying frequency response .

2. The two-stream Mamba method with propagation environment constraints for multi-input multi-output channel prediction of maritime unmanned aerial vehicles according to claim 1, characterized in that, Step S2 further includes: The Pierson-Moskowitz power spectrum combined with the Beaufort wind classification was used to divide the sea surface environment into several typical sea state levels; Under each sea state level, different drone receiver movement speeds and communication signal-to-noise ratio conditions are set to form a simulation configuration with multiple scenarios and conditions. Based on the low-altitude multi-input multi-output channel model of the sea surface constructed in step S1, channel simulations are carried out under various simulation configurations. High-dimensional time series data containing scene topology, environmental parameters and channel state information are generated through simulation, and a high-fidelity sea surface channel dataset covering multiple sea states and multiple motion states is constructed.

3. The two-stream Mamba method with propagation environment constraints for multi-input multi-output channel prediction of maritime unmanned aerial vehicles according to claim 1, characterized in that, Step S3 further includes: The continuous historical observation sequences in the high-dimensional time series data obtained in step S2 are discretized. By combining the core physical parameters affecting the marine electromagnetic propagation environment, a mapping relationship between the external meteorological environment and the internal channel dynamic response is established, and the environmental feature vector at time t is constructed. ;in, For wind speed, For the effective wave height, The height of the evaporation waveguide. This is the sea surface roughness coefficient. The distance between the transmitting and receiving ends, These are the yaw, pitch, and roll angles of the transmitter. For the receiver's yaw angle, pitch angle, and roll angle; The MIMO system at time t Mapping the complex channel tensor of each antenna pair to the real domain, we construct the space-time-frequency channel state information vector at time t as follows: ;in, For transceiver antenna pairs The time-varying frequency response between; By combining the environmental feature vector and the channel state information vector, two independent and parallel input streams, namely the environmental feature stream and the channel state information stream, are constructed to generate the input tensor of the prediction model. The environmental feature stream and the channel state information stream are respectively composed of the environmental feature vector and the channel state information vector at the corresponding time point in a time sequence.

4. The two-stream Mamba method with propagation environment constraints for multi-input multi-output channel prediction of maritime unmanned aerial vehicles according to claim 1, characterized in that, In step S4, the input sequences of the environmental feature stream and the channel state information stream are processed in batches and mapped to the high-dimensional latent space through independent linear projection layers to obtain two sets of corresponding embedding representations, denoted as environmental feature embedding and channel state embedding, respectively; where the dimension of each set of embedding representations is B×L×D, B is the batch size, L is the sequence length, and D is the hidden dimension.

5. The two-stream Mamba method with propagation environment constraints for multi-input multi-output channel prediction of maritime unmanned aerial vehicles according to claim 1, characterized in that, In step S5, the prediction model sequentially includes a lexicalization layer, a dual-stream Mamba coding layer, a cross-stream gating unit layer, and a linear decoding layer; The lexicalization layer performs lexicalization on the two sets of consecutive embedding representations obtained in step S4 to obtain the environmental feature stream lexical sequence. and channel state information stream lexical sequence Among them, environmental feature flow word sequence Temporal characteristics used to characterize environmental factors including sea surface weather and UAV attitude; channel state information stream term sequence Used to characterize the dynamic changes in the space-time-frequency response of a MIMO channel; The dual-stream Mamba coding layer includes a slow branch and a fast branch; wherein, the slow branch adopts a Bi-Mamba structure to encode the context feature stream token sequence. Temporal modeling is performed to extract the long-term evolution trends of environmental parameters, including sea surface wind speed, wave height, evaporation waveguide height, and UAV attitude, to characterize the continuous changes in the sea surface propagation environment and form environmental constraint features with physical prior attributes. The fast branch uses a Bi-Mamba structure to stream the term sequence of channel state information. Time-series modeling is performed to capture the details of instantaneous channel abrupt changes caused by multipath superposition, Doppler jitter, and phase delay fluctuations, thus characterizing the short-time rapid dynamic response of non-stationary sea surface channels and forming detailed channel state features. ; The cross-current gating unit layer will define the environmental constraint characteristics of the slow branch output. A dynamic gating mask is generated through a linear layer, and the mask is fine-tuned element-by-element based on the physical boundary of sea surface radio wave propagation: the mask value is increased for channel mutation dimensions consistent with the long-term evolution trend of the environment to retain information on compliant environmental mutations; the mask value is decreased for noise fluctuation dimensions that deviate from physical laws to suppress invalid anomalous components; and this mask is used to extract detailed channel state features from the fast tributary output. Modulation is performed, and then the two-stream features are fused through residual connection and normalization operations to obtain the fused features. This allows the channel characteristics to carry constraint information about the sea surface propagation environment; The linear decoding layer will input fused features After time-domain extrapolation using layer normalization and feedforward neural network, the data is then restored to the original channel physical observation space through a linear projection layer, outputting the channel state prediction results for the MIMO of the UAV at future time.

6. The two-stream Mamba method with propagation environment constraints for multi-input multi-output channel prediction of maritime unmanned aerial vehicles according to claim 5, characterized in that, The cross-flow gating unit layer will first carry the environmental constraints characteristics of the long-range evolution law of the sea surface environment. The input is fed into a linear transform layer for feature dimension alignment. A dynamic gating mask is generated through the linear layer and the Sigmoid activation function. This mask adaptively adjusts the weight allocation of channel features based on the changing trends of environmental parameters, including sea surface wind speed, wave height, evaporation waveguide height, and UAV attitude. Then, the dynamic gating mask is used to refine the channel state details of the fast branch output, which include multipath superposition, Doppler jitter, and phase delay fluctuations. Element-wise Hadamard product modulation is performed to obtain detailed channel state features. The various dimensions are adaptively weighted to suppress channel noise and non-stationary fluctuations that do not conform to the physical laws of the environment; finally, the modulated channel features are compared with the original channel state details. Residual connections are performed, and then layer normalization is used to complete the fusion of dual-stream features, resulting in fused features that take into account both long-term environmental constraints and instantaneous channel details.

7. The two-stream Mamba method with propagation environment constraints for multi-input multi-output channel prediction of maritime unmanned aerial vehicles according to claim 6, characterized in that, The process of adaptively adjusting the weight allocation of channel features based on the changing trends of environmental parameters, including sea surface wind speed, wave height, evaporation waveguide height, and UAV attitude, includes the following steps: Environmental constraints that carry the long-term evolution law of the marine environment The input linear transform layer is used for dimension alignment to generate detailed features of the fast tributary channel state. Dimensional matching mask generation features; A dynamic gated mask is generated by using a linear layer and a Sigmoid activation function. The mask has a value range of [0,1], and each mask element corresponds to one dimension of the channel state detail features. The pre-trained physical agent network is invoked, and environmental parameters including sea surface wind speed, wave height, evaporation waveguide height, and UAV attitude are used as input to output the dynamic physical boundary corresponding to the propagation of sea surface radio waves in real time. The dynamic physical boundary includes the multipath delay interval, the maximum Doppler frequency shift threshold, and the upper and lower limits of the signal energy envelope. Based on the dynamic physical boundary, the channel state details are described. Compliance is determined on a dimensional basis. Dimensions whose feature values ​​fall within the physical boundary range are determined to be compliant channel dimensions, while dimensions whose feature values ​​exceed the physical boundary or exhibit irregular instantaneous jumps are determined to be noise dimensions. Based on the dimensionality determination results, the dynamic gating mask is fine-tuned element by element: the mask elements corresponding to the compliant channel dimension are numerically increased to amplify the effective feature weights, while the mask elements corresponding to the noise dimension are numerically decreased to weaken the abnormal fluctuation weights, and the mask values ​​remain within the [0,1] range after correction; then the fine-tuned mask is used to analyze the channel state detail features. Element-wise Hadamard product modulation is performed to complete the weight allocation of each channel dimension, thereby suppressing the portion of channel noise and non-stationary fluctuations that does not conform to the physical laws of the environment.

8. The two-stream Mamba method with propagation environment constraints for multi-input multi-output channel prediction of maritime unmanned aerial vehicles according to claim 1, characterized in that, In step S5, the prediction model is trained and its parameters are updated using a joint loss function based on physical environment boundary constraints. This joint loss function is a weighted sum of the channel prediction mean square error term and the physical environment boundary constraint regularization term. ； in, For the total loss, This is the balance coefficient; This is the channel prediction mean square error term, used to constrain the deviation between the channel state prediction results output by the model and the actual channel state; The physical environment boundary constraint regularization term utilizes a pre-trained physical proxy network and channel feature extraction operators to construct dynamic physical boundary conditions, penalizing abnormal results in the channel state information flow prediction results that violate the physical constraints of sea surface channel propagation.

9. The two-stream Mamba method with propagation environment constraints for multi-input multi-output channel prediction of maritime unmanned aerial vehicles according to claim 8, characterized in that, The physical agent network is a multilayer perceptron architecture, and pre-training is completed through offline supervised learning before training the prediction model. In step S5, the physical agent network takes the environmental feature stream as input and infers a multi-dimensional physical evolution boundary threshold vector, which includes the upper bound of the maximum Doppler frequency shift, the maximum multipath delay spread limit, and the energy envelope limit.

Images